Research On Wideband DOA Estimation Method Based On Bayesian Blind Separation

Wideband DOA(Direction of arrival)estimation is a key technology of acoustic imaging.In recent years,the academic community has explored a variety of DOA estimation methods.Among them,DOA estimation based on Bayesian blind source separation has become one of the research hotspots in academia.However,its application cases under strong noise and reverberant environments are rare and cannot meet the requirements of high-precision mobile acoustic imaging.In order to meet the needs of practical application,this paper studies the DOA estimation method of wideband time-varying multi-source based on Bayesian blind source separation in the strong noise and reverberation environment.The main work is as follows.1.A wideband DOA estimation method based on Gaussian process regression-unscented Kalman filter-nonnegative matrix factorization deconvolution(GPR-UKF-NMFD)is designed and realized.GPR is first used to establish a state-space model,combined with UKF to achieve real-time state estimation,and then NMFD is used to complete the blind separation and dereverberation of the mixed signal,and generalized cross correlation phase transformation(GCC-PHAT)to pre-estimate value of the signal source position.Finally,the three-dimensional spatial spectrum of the signal source is obtained based on the minimum variance distortionless response(MVDR)beamformer.2.A wideband DOA estimation method based on the variational Bayesian blind separation is studied.The local complex Gaussian model and the non-negative matrix factorization probability frame are used as the priors of the source signal,and the time-varying mixed matrix is simulated by a continuous-time random process,as well as the multi-channel mapping of the observation signal is realized by using set empirical mode decomposition principal component analysis.On this basis,the variational Bayesian expectation maximization is used to obtain the separated signals.Finally,the DOA estimation of the signal source is realized by GCC-PHAT and MVDR beamformer.3.A wideband DOA estimation method based on time-varying online blind separation method is analyzed and implemented.The complex fast independent component analysis and the second-order blind identification algorithms are used to obtain the pre-estimated values of the signal source and the mixed system.The GPR model is used to learn the time structure of the source signal and the mixed matrix.Model parameters and hyperparameters are obtained through variational Bayesian inference,and the hyperparameters are passed as a prior for the next frame,thus realizing online Bayesian learning.Then,the DOA and spatial spectrum of the signal source are obtained by GCC-PHAT and MVDR beamformer.4.Based on the research of wideband DOA estimation methods in the previous chapters and inspired by convolutional neural networks,an unsupervised deep blind beamforming algorithm is proposed.This algorithm is transplanted into the acoustic camera system,which realizes the time-varying blind separation,localization and dereverberation of the measured multiple sound sources.Through the analysis and verification of computer simulation and hardware in the loop simulation experiment,the experimental results show that the wideband DOA estimation method proposed in this paper can meet the requirements of practical application.